MemoVAD: Resource-Efficient Video Anomaly Detection via Dynamic Semantic Memory in Edge Computing Scenarios

Researchers introduce MemoVAD, an edge-cloud collaborative framework that enables efficient video anomaly detection on resource-constrained devices by selectively querying cloud-based Vision-Language Models only for uncertain or novel scenarios. The system uses dynamic semantic memory to cache verified patterns, reducing computational overhead while maintaining detection accuracy on surveillance tasks.

MemoVAD addresses a critical constraint in deploying advanced AI systems at the network edge: the collision between computational sophistication and hardware limitations. Traditional video anomaly detection requires either compromising accuracy with lightweight models or accepting prohibitive latency from cloud-dependent processing. This research demonstrates a pragmatic hybrid approach that leverages the complementary strengths of edge and cloud resources.

The technical innovation centers on intelligent gatekeeping rather than brute-force optimization. By using Uncertainty-Aware Gating grounded in Subjective Logic, the system identifies which video clips genuinely require Vision-Language Model consultation, dramatically reducing communication costs and cloud API calls. The Dynamic Semantic Memory acts as a learned cache layer, storing verified anomaly prototypes for rapid pattern matching without repeated cloud queries. This design mirrors real-world surveillance workflows where most frames contain expected patterns but occasional anomalies demand higher-level semantic reasoning.

For edge computing infrastructure operators and surveillance vendors, MemoVAD presents meaningful efficiency gains. Reduced cloud bandwidth consumption directly translates to lower operational costs, while maintaining competitive accuracy demonstrates that resource constraints need not mandate accuracy sacrifices. The framework's validation on UCF-Crime and XD-Violence datasets using actual edge hardware—not simulations—strengthens credibility for production deployment.

The broader implication extends beyond surveillance. As organizations increasingly deploy AI at network edges for real-time processing, selective cloud augmentation patterns will likely proliferate across computer vision, anomaly detection, and other latency-sensitive applications. Future development should explore transferability across different edge hardware classes and investigation of privacy implications when edge devices cache sensitive visual features.

• →MemoVAD reduces cloud communication overhead by querying Vision-Language Models only for high-uncertainty scenarios, improving efficiency without sacrificing accuracy
• →Dynamic Semantic Memory caches VLM-verified anomaly prototypes, enabling edge devices to progressively incorporate advanced semantics through retrieval rather than computation
• →Real edge device experiments on surveillance datasets demonstrate substantial performance gains over state-of-the-art approaches
• →The framework addresses the fundamental constraint between semantic richness requirements and limited edge device computational resources
• →Uncertainty-Aware Gating grounded in Subjective Logic provides principled decision-making for selective cloud offloading